Lecture 5: Interval Estimation

25
Machine Learning for Language Technology 2015 h6p://stp.lingfil.uu.se/~san?nim/ml/2015/ml4lt_2015.htm Sta%s%cal Inference (2) Interval Es?ma?on Marina San%ni san%[email protected]fil.uu.se Department of Linguis%cs and Philology Uppsala University, Uppsala, Sweden Autumn 2015

Transcript of Lecture 5: Interval Estimation

Page 1: Lecture 5: Interval Estimation

Machine  Learning  for  Language  Technology  2015  h6p://stp.lingfil.uu.se/~san?nim/ml/2015/ml4lt_2015.htm  

   Sta%s%cal  Inference  (2)  

Interval  Es?ma?on  Marina  San%ni  

 san%[email protected]  

 Department  of  Linguis%cs  and  Philology  Uppsala  University,  Uppsala,  Sweden  

 Autumn  2015  

   

Page 2: Lecture 5: Interval Estimation

Acknowledgements  

•  The  web,  sta%s%cal  websites,  online  calculators    

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 2

Page 3: Lecture 5: Interval Estimation

Outline  

•  Confidence  intervals  – On  propor%ons  – On  means  

•  Standard  error  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 3

Page 4: Lecture 5: Interval Estimation

Sta%s%cal  Inference:    Interval  Es%ma%on  

•  Suppose  we  measure  the  error  of  a  classifier  on  a  test  set  and  obtain  a  certain  numerical  error  rate,  eg.  25%.    

•  This  corresponds  to  a  success  rate  of  75%.    •  This  is  an  es%mate  on  a  sample  (our  dataset).    

•  What  can  we  say  about  the  "true"  success  rate  on  the  target  popula%on?    

•  Remember:  We  have  observed  the  propor%on  of  correct  classifica%ons  on  a  sample,  while  the  popula%on  is  unknown  to  us.  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 4

Page 5: Lecture 5: Interval Estimation

Our  prac%cal  ques%on  is…  

l  When the estimated success rate is 75%, how close is this value to the true success rate, ie the success rate on the population?

♦ Depends on the amount of sample size

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 5

Page 6: Lecture 5: Interval Estimation

What  is  a  confidence  interval?    •  In  sta%s%cal  inference,  one  wishes  to  es%mate  popula%on  

parameters  using  observed  sample  data  

•  Confidence  intervals  provide  an  essen%al  understanding  of  how  much  faith  we  can  have  in  our  sample  es%mates  

•  A  confidence  interval  is  a  range  computed  using  sample  sta%s%cs  to  es%mate  an  unknown  popula%on  parameter  with  a  given  level  of  confidence.    

–  For  example,  we  want  to  say:  “we  are  80%  certain  that  true  popula%on  propor%on  falls  within  the  range  of  73.25%  and  76.75%  

–  We  usually  write  the  confidence  interval  in  this  way:  [0.732,0.767]        

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 6

Page 7: Lecture 5: Interval Estimation

Generally  speaking...  

•  A  confidence  interval  is  constructed  by  taking  the  point  es%mate  (p̂)  plus  and  minus  the  margin  of  error.    

•  The  margin  of  error  is  computed  by  mul%plying  a  z  mul%plier  by  the  standard  error,  SE(p̂).    

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 7

Page 8: Lecture 5: Interval Estimation

Defini%on:  Standard  Error          •  Standard  error  is  a  sta%s%cal  term  that  measures  the  accuracy  with  which  a  sample  represents  a  popula%on.    

•  In  sta%s%cs,  a  sample  mean  or  a  sample  propor%on  deviates  from  the  actual  mean  or  propor%on  of  a  popula%on;  this  devia%on  is  the  standard  error.    The  smaller  the  standard  error,  the  more  representa%ve  the  sample  will  be  of  the  overall  popula%on.  The  standard  error  is  also  inversely  propor%onal  to  the  sample  size;  the  larger  the  sample  size,  the  smaller  the  standard  error  because  the  sta%s%c  will  approach  the  actual  value.    

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 8

Page 9: Lecture 5: Interval Estimation

The  Mul%plier  The multiplier is a constant that indicates the number of standard deviations in a normal curve. The larger the multiplier, the higher the confidence level, the narrower the confidence interval, the more reliable the prediction of the performace.The constant for 80% percent confidence intervals is 1.28 (see table or use a calculator: http://www.gngroup.com/stat.html )

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 9

Page 10: Lecture 5: Interval Estimation

Confidence  intervals  

•  Confidence  intervals  of  a  propor%on  •  Confidence  intervals  of  the  mean  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 10

Page 11: Lecture 5: Interval Estimation

Confidence  interval  for  propor%on  

•  A  confidence  interval  for  a  propor%on  is  constructed  by  taking  the  point  es%mate  (p)̂  plus  and  minus  the  margin  of  error.  The  margin  of  error  is  computed  by  mul%plying  a  mul%plier  by  the  standard  error,  SE(pˆ).  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 11

Page 12: Lecture 5: Interval Estimation

The  standard  error  of  propor%on:  p̂  (p-­‐hat)  

•  The  standard  error  is  an  es%mate  of  the  standard  devia%on  of  a  sta%s%c.    

•  This  is  the  formula  of  the  Standard  Error  of  an  es%mated  propor%on  (the  hat  always  represents  an  es%mate)  

•  p̂  =  es%mated  propor%on  •  n  =  sample  (number  of  observa%ons)  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 12

Page 13: Lecture 5: Interval Estimation

Our  prac%cal  ques%on  is…  

l  When the estimated success rate is 75%, how close is this value to the true success rate, ie the success rate on the population?

♦ Depends on the amount of sample size

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 13

Page 14: Lecture 5: Interval Estimation

Confidence  intervals  on  our  propor%on  

l  We can say that our point estimate 75% lies within a certain specified interval with a certain specified confidence (say 80%):

l  Example: S=750 successes in N=1000 trials l  Estimated success rate: 75% l  How close is this to true success rate p?

l  Answer: with 80% confidence p in [73.2,76.7] l  Another example: S=75 and N=100

l  Estimated success rate: 75% l  Answer: With 80% confidence p in [69.1,80.1]

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 14

Page 15: Lecture 5: Interval Estimation

l  p ̂= 75%, n = 1000, confidence = 80% (so that z = 1.28):

p∈[0.732,0.767]

l  p ̂= 75%, n = 100, confidence = 80% (so that z = 1.28): p∈[0.691,0.801]

l  Usually the normal distribution assumption is only valid

for large n (i.e. n > 100) l  In a case like this: p ̂= 75%, n = 10, confidence = 80%

(so that z = 1.28): p∈[0.549,0.881]

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 15

Page 16: Lecture 5: Interval Estimation

Confidence  Interval  Calculator  for  Propor%ons  hdps://www.mccallum-­‐layton.co.uk/tools/sta%s%c-­‐calculators/confidence-­‐interval-­‐for-­‐propor%ons-­‐calculator/    

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 16

Page 17: Lecture 5: Interval Estimation

Confidence  intervals  around  the  mean  

Confidence  intervals  are  calculated  based  on  the  standard  error  of  the  mean  (SEM):    s  =  sample  standard  devia%on  (see  formula  below)    n  =  sample  (number  of  observa%ons)    The  following  is  the  sample  standard  devia%on  formula  (see  also  lecture  2):  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 17

Page 18: Lecture 5: Interval Estimation

Example:  How  to  compute  the  confidence  interval  of  teh  mean  

 A  brand  ra%ng  on  a  five  point  scale  from  62  par%cipants  was  4.32  with  a  standard  devia%on  of  .845.  What  is  the  95%  confidence  interval?    1)  Find  the  mean:  4.32  2)  Compute  the  standard  devia%on:  .845  3)  Compute  the  standard  error  by  dividing  the  standard  devia%on  by  the  square  root  of  the  sample  size:    .845/  √(62)  =  .11  4)  Compute  the  margin  of  error  by  mul%plying  the  standard  error  by  2  (it  is  common  to  round  up  1.96  to  2).  =  .11  x  2  =  .22  5)  Compute  the  confidence  interval  by  adding  the  margin  of  error  to  the  mean  from  Step  1  and  then  subtrac%ng  the  margin  of  error  from  the  mean:    

     Lower  limit:  4.32-­‐.22  =  4.10    Upper  limit:  4.32+.22  =  4.54      

The  95%  confidence  interval  is  4.10  to  4.54.  We  don't  have  any  historical  data  using  this  5-­‐point  branding  scale,  however,  historically,  scores  above  80%  of  the  maximum  value  tend  to  be  above  average  (4  out  of  5  on  a  5  point  scale).    Therefore  we  can  be  fairly  confident  that  the  brand  is  at  least  above  the  average  threshold  of  4  because  the  lower  end  of  the  confidence  interval  exceeds  4.    Source:  hdp://www.measuringu.com/blog/ci-­‐five-­‐steps.php    

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 18

Page 19: Lecture 5: Interval Estimation

Confidence  Interval  Calculator  for  Means    

hdps://www.mccallum-­‐layton.co.uk/tools/sta%s%c-­‐calculators/confidence-­‐interval-­‐for-­‐mean-­‐calculator/        

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 19

Page 20: Lecture 5: Interval Estimation

Quiz  1:  Confidence  Interval  (Mean)  You  take  a  sample  of  25  test  scores  from  a  popula%on.  The  sample  mean  is  38  and  the  populaton  standard  devia%on  is  6.5.  What  is  the  95%  confidence  interval  of  the  mean?    1.  [37.49,38.51]  2.  [36.49,39.51]  3.  [35.45,40.55]  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 20

Page 21: Lecture 5: Interval Estimation

Calculator    hdps://www.mccallum-­‐layton.co.uk/tools/sta%s%c-­‐calculators/confidence-­‐

interval-­‐for-­‐mean-­‐calculator  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 21

Page 22: Lecture 5: Interval Estimation

Quiz  2:  Confidence  Interval  (Propor%on)  

747  out  of  1168  female  students  said  they  always  use  a  seatbelt  when  driving.  What  is  the  99%  confidence  interval  for  the  propor%on  of  female  students  in  the  popula%on  who  always  use  a  seatbelt  when  driving?  1.  [.612,.668]  2.  [.604,.676]  3.  None  of  the  above  

 

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 22

Page 23: Lecture 5: Interval Estimation

Calculator  hdps://www.mccallum-­‐layton.co.uk/tools/sta%s%c-­‐calculators/confidence-­‐

interval-­‐for-­‐propor%ons-­‐calculator/      

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 23

Page 24: Lecture 5: Interval Estimation

Conclusions  •  A  confidence  interval  is  a  range  of  values  that  is  likely  to  contain  an  

unknown  popula%on  parameter.    

•  Confidence  intervals  serve  as  good  es%mates  of  the  popula%on  parameter  because  the  procedure  tends  to  produce  intervals  that  contain  the  parameter.    

•  Confidence  intervals  are  comprised  of  the  point  es%mate  (the  most  likely  value)  and  a  margin  of  error  around  that  point  es%mate.  The  margin  of  error  indicates  the  amount  of  uncertainty  that  surrounds  the  sample  es%mate  of  the  popula%on  parameter.    We  will  resume  this  topic  in  Lecture  8.  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 24

Page 25: Lecture 5: Interval Estimation

The  end  

Lecture  5:  Statistical  Inference  2:  Interval  Estimation 25